Discussion
"snippets from LLM responses were copied as a starting point to be modified into game code—we’d estimate that well under 0.1% of Big Walk’s code had anything to do with this usage, and generative AI was not used in any other context"
I'm against use of LLMs in programming (and most contexts in general), but canceling a game over the above scenario is not a level of extremism I can join.
IMO a straightforward non-extremist line to draw is whether agents have access to a codebase or not.
@optozorax @runevision replace "LLM" here with something you consider to be abhorrent, see what it does to how this reads. "A slave" or "a child labourer" might work. Not saying LLMs are equivalent, but it should unlock a better ability to see the other side.
@mdiluz do you mean that LLM might have consciousness or qualia/feelings? That's a valid argument and in the future it will become much more important and relevant. It's good to hear such nuanced take. But right now I don't think that using LLM to write code isn't as bad as having "a human slave" write this code. If you don't use recently found "torture vector", of course. Using it on LLMs is beyond unethical.
@optozorax no, I literally said "Not saying LLMs are equivalent". My point is if you replace "LLM" in your own statements with something that *you* find problematic to use, you'll see why others have a problem.
Other's see LLMs themselves as problematic on about 15 levels, so arguing "but they work" is irrelevant. It'd be the same kind of fallacy as saying "but my machine that kills 50 whales for every line of code takes only 1h instead of 40h to write code, is that bad usage of that machine?"
@runevision
> I'm against use of LLMs in programming
Why? For example, what if I made my program, then I made a lot of tests and ask LLM agent to optimize it, hill climb some measurement that I provided, it does some intricate things that doesn't change logic, but do write (and discard!) a lot of code that actually works faster (let's say order of magnitude faster). And it takes 1h of my time instead of 40h. Is this bad usage of LLM?